Project Title Description Hardware Software Electrical component themed AI detection and identification. A mounted camera above a surface (part of the product). Produces a controlled environment live feed for the application. An application running inference on a live USB camera feed (optionally imported picture or video) Application Modification of the provided data to simulate differences in the environment that would occur in a real life scenario, and to provide a challenge in form of imperfections to both train against, and test against. Augmentation examples Addition of glare Rotation Blurring Addition of spots GUI The GUI Application is based on Qt Creator, using C++ Inference Running on C++ Utilising Summary Detection of objects from an image via Inference Detect and display boundaries for each identified class, and the confidence value of this detection from the input image using Inference. Identification Post-processing of the components in the bounding boxes detected by inference, which may have additional information that can be identified by a variety of approaches. Examples LEDs Resistors Resistor code value LED color Technology AI based Electronic Component Identifier IC Components Pin count Information written on the component Features Inference Classes present in the dataset that the model will be trained upon Resistor Diode Capacitor LEDs Integrated Circuits AC DC LDR Milestones Base camera rig Initial inference model training Inference running Testing with video footage from a mobile device Research Pre-Trained Models Ultralytics YOLO Focus Audience Set Rig The set position of the camera, a significant reduction in distance between the objects, and significant consistency of the lighting provided by the ring light, and the static background - will boost the confidence of the inference considerably. Training Running Post-processing Rationale Timeline Gantt Chart Live training YOLOv5 Training 1st batch, test run Image Count 100 Classes (1 Total) Resistor 2nd batch Image Count Training Evaluation 20 Training Evaluation 1800 540 3rd batch Image Count Training Evaluation 2393 724 Classes (9 Total) red_led green_led blue_led yellow_led ac_capacitor dc_capacitor resistor sip_resistor pcb_terminal Classes (10 Total) red_led green_led blue_led yellow_led ac_capacitor dc_capacitor resistor sip_resistor pcb_terminal metal_nut Augmentation Default Augmentation Default Average time per epoch 34 seconds Epoch count 400 Epoch count 250 Epoch count 300 Augmentation Default YOLOv8 Due to the angle and lighting both being known and mostly set thanks to using a set rig, the input dataset does not need to cover angles and lighting outside what the rig will expose it to during runtime. The sum of all the points covered above results in a significant reduction in data required to train when compared to a setup without a set rig, for equivalent confidence values during runtime. The angle range is reduced only to looking from top to down, eliminating the rest of the angle range. While the lighting will change depending on the room conditions, the ring light around the camera will provide significant consistency in lighting. While this does not not eliminate the necessity to train against various lighting conditions, it does reduce their significance and increase certainty of the detection. Only the components being detected need to be trained in all angles, as opposed to the camera gathering the dataset requiring to be positioned in different angles. Having a top to down view also eliminates the majority of issues that come with glare from high luminosity bodies, such as clouds or the sun. A set rig significantly limits the distance that the objects will be from the camera during runtime, allowing for further confidence in the predictions. Static background Angle range Lighting Apart from dust or unexpected objects present on the rig's surface, which should be removed before usage - the background that the objects are in front of will stay mostly consistent. This reduces the necessity to gather data of the same object under backgrounds that are not expected to be used during runtime. While this project may be retrained and refocused to be utilised for many different fields - it is trained for electrical component identification, which is focused towards engineers. Architectures This project focuses on both existing engineers, and ones that are interested in becoming engineers. Having access to the provided by the project quick identification of components, count of each, and any potential additional information saves time spent manually analysing this information. Average time per epoch 2 minutes Average time per epoch 2 minutes and 20 seconds SIP Resistor Singular Acronyms SIP - Single Inline Package Introduction GPU - Graphics Processing Unit CPU - Central Processing Unit AI - Artificial Intelligence LDR - Light Dependant Resistor LED - Light Emitting Diode AC - Alternating Current DC - Direct Current PCB terminal PCB - Printed Circuit Board The most prominent color may be identified by sorting all the colors from the image into their hue values, and checking which hue is most active. The color codes can be identified by processing the image using filters and otherwise until only the prominent colors remain. These can be processed into the actual ohm value. Then, the positions of the color codes relative to the body of the resistor can be used to identify the specific positions and order of the color codes. The pin count can be identified by processing the image using filters until there is a clear contrast between the body of the chip, and the pins. One approach that could help identify the number of pins would be drawing a line between two of the pins and seeing how many of the pins touch this line. Taking the line that touches the most pins would provide the pin count of this IC. OCR may be used to extract the text based information. OCR - Optical Character Recognition Software based reading alphabetical characters from an image that contains written text. Input image Inference method Algorithm method Different color LEDs may be trained as individual classes. Has the disadvantage of requiring training for each individual LED separately, as opposed to one generic LED. Has the advantage of working on any LED. Raw input High contrast filter Colors histogram Approaches after filtering Results show clearly prominent yellow, which is accurate. Has the disadvantage of potentially giving false information if the background is too vibrant. Contrast approach HSV Taking the average of all the pixels hue values that have a value above a certain threshold. Around 0.7 on a range from 0 to 1 should be appropriate. Hue is in the range of 0 to 360 degrees. The pink dots represent the pixel values obtained from the previous step. Taking the average of this data, the result would land in the degree value that can be easily determined as yellow, by separating the hue circle into sections of colors by degrees ranges. HSV, or Hue Saturation Value, is an alternative way to represent colors. It can be advantageous over RGB in situations such as this. RGB - Red Green Blue Commonly used to referred to a way of defining colors by their Red, Green and Blue properties. HSV - Hue Saturation Value Commonly used to referred to a way of defining colors by their Hue, Saturation and Value properties. Yellow is between 72° and 108° degrees on the hue circle. Note: This example would ignore colors that are darker than 0.7, on a range of 0 to 1. Sorted from highest priority, to lowest. Setting up the camera on a rig. Base GUI GUI with essentials to interface the camera through USB, with A live display from the camera on the rig. Ability to export images by pressing a button. Support for running Inference. ~100 images of a single class, taken from the rig for initial training and testing of the model. Initial dataset gathering For the purpose of testing inference on the rig. Proof of concept. The results will not be perfect as the dataset is minimal, and only contains 1 class. Further dataset gathering At least 250 pictures of each class of every component that the project is designed to detect. Further model training This training will take considerably longer than the initial training. Around 2 minutes per epoch, and should be ran for at least 300 epochs. The initial training should not take long at all, and does not require to be polished. Training for ~100 epochs should be sufficient, with each epoch taking ~20 seconds on the machine available. Rig Model Training via Deep Learning Machine used Personal Computer CPU GPU AMD RyzenTM 7 5800X3D Core count 8 Base clock frequency 3.4GHz L3 Cache 96MB Maximum operating temperature 90°C Thread count 16 GeForce RTX 3060 Ti Memory 8192MB CUDA core count 4864 Capacity Type GDDR6X Base clock frequency 1.41GHz The goal is to reach 0.8 from range of 0 to 1 confidence values. Ability to gather further information from the detection bounding boxes provided by the inference. After the previous steps are in good shape, investigation of moving the inference to a mobile device will begin. If the confidence values are not up to standard, more data will be gathered from this and potentially other mobile devices, and further training will follow, until the results are adequate. If adequate results are achieved before the deadline of this project, deployment to a mobile device will be started. If the frame rates are not sufficient enough, the inference may be ran on still images to improve user experience. Optional: Ability to label the images from the device, without requiring external software. It may be advantageous given the timeframe of the project to instead gather data during a session and labelling it afterwards. Memory Capacity Type 2x16GB DDR4 Frequency 3.6GHz Brand Corsair Name Vengeance RGB PRO SL Link https://www.corsair.com/ww/en/Categories/Products/Memory/Vengeance-RGB- PRO-SL-White/p/CMH16GX4M2E3200C16W Brand AMD Name Ryzen 7 58700X3D Link https://www.amd.com/en/products/cpu/amd-ryzen-7-5800x3d Brand NVIDIA Series 30 Name RTX 3060Ti Link https://www.nvidia.com/en-gb/geforce/graphics-cards/30-series/rtx-3060-3060ti/ CUDA Special cores that are designed for compute-intensive tasks. These run parallel with the CPU, and may also run parallel with multiple GPUs. They are perfect for deep learning, as deep learning is incredibly compute intensive. Deep learning training times are predictable, and stay mostly constant between epochs. This means that there are no race conditions, and the more processing power available, the quicker the epoch will finish. Each of these steps should be polished before continuing to the next one, to provide a solid foundation for the next step to be based on. Analysis of the models Brief History YOLO, which stands for You Only Look Once is a popular image segmentation and object detection model that was originally developed by Joseph Redmon and Ali Farhadi. The first version was released in 2015, and it very quickly became popular due to the significantly superior speed and accuracy when compared to other architectures. YOLOv1 YOLOv4 Released in 2018, Introducting of Mosaic data augmentation, and a new and improved loss function - decreasing time taken to achieve better results for the trained model. YOLOv5 Released in 2020, Introducing support for Object Tracking - which allows following a moving object, and Panoptic Segmentation, which allows identification of overlapping objects, with accurate bounding boxes. Ultralytics YOLOv8 The latest version of YOLO as of today. YOLOv8 is a state-of-the-art model that builds upon the already very successful previous YOLO versions, introducing new performance and flexibility features. Full support for previous YOLO versions, making it incredibly convenient for existing users of previous YOLO versions to take advantage of the new features. Versions Comparison In general, YOLOv8 is superior to all of its predecessors. While YOLOv5 is mostly underperforming when compared to the next versions, it is important to note how incredibly minimal the delays are even on a version so outdated now. YOLO offers pre-trained models that are used to start train custom models. Each model has its advantages and disadvantages, and should be picked depending on the project. Size mAP single-model single-scale values while detecting on the COCO val2017 dataset. Speed Averaged time taken using the Amazon EC2 P4d instance on the COCO dataset. The pixel height and width the model operates up to. Params (In Millions) The number of parameters that are tweaked per epoch while training, and processed during inference. FLOPS Floating Point Operations Per Second A measure based on Floating Point Operations that is relevant in the field of Deep Learning. Diminishing results can be observed on the mAP values when compared to the time taken (Speed). Model properties In some circumstances, maximum precision is essential, and is prioritised over the hardware requirements. This is when a higher model should be chosen. In the scope of this project - several pre-trained models have been used, including both YOLOv5 and YOLOv8, for the purpose of comparison. In comparison of YOLOv5 and YOLOv8 versions - a clear advantage can be seen when taking into account the size of the model (param count), and the resulting mAP output, as well as the time taken. Architecture choice YOLO has been chosen as the architecture that this project utilises for the AI detection. At the start of the project, there was already a high bias towards YOLO due to the highly positive personal past experience with YOLOv5 and all the incredible features that it offers. Upon release of YOLOv8 and all the superior features and specifications that it provides on top of the previous versions which this project may expand into - the YOLO family was an obvious choice in the architecture that will be used for the project. As the name suggests, YOLO focuses on detection of multiple classes in a single "look", which is a single analysis of the entire input image. When compared to many other architectures before YOLO, realistically, no matter how quick the other architectures may be - this is an incredibly superior approach, as other architectures would approach detection by reanalysing the entire image for every single class that the model was trained for - increasing the time taken per detection additively per class. An approach like this may seem too good to be true, and that it should come with signficant cost to the speed and confidence of the model. But when the results are analysed - that could barely be further from the truth. YOLO is an incredibly efficient and accurate architecture. These days most sophisticated architectures approach object detection similarly to YOLO, but YOLO is still a state-of-the-art architecture that continues to improve and grow to this day. Internal AI Object Detection steps Classification Object Detection Segmentation The process of identifying the exact bounding box of the item detected. The Bounding by a box of the classified segments of the image. The identification of a part of an image believe to contain an item of a class the model was trained to detect. Visual examples Resizing Joining up of multiple images to create new ones The reduction of data required to train makes it feasible to train relatively high quality models from data gathered and trained from home. Marking Codes Hardware Raspberry Pi 4 Beaglebone Nvidia Jetson Nano Intel Neural Compute Stick 2 Specifications Processor Base Frequency 700MHz Memory 2GB Specifications Core Count (SHAVE) 16 Advantage Offers computational power through a USB connection - can be used to run Inference on existing devices, such as a laptop. Specifications Specifications Core Count 4 Maximum Frequency 700MHz Resistors and Inductors Capacitors ICs Color coded Number coded Android Phone Specifications depend on the specific device. Widely and easily accessible. The vast majority of mobile phones on the market today have a built-in camera. YOLO - You Only Look Once An image detection architecture that the project is based on. CUDA cores provided by the GPU CPU Inference Training Personal Computer Rented Dedicated Server Advantages Disadvantages Advantages Disadvantages Local - Provided a local machine is already owned, it is immediately available. Utilises multiple GPUs - Quicker epoch computations, resulting in quicker training. Cloud based Allows for parallel computing, as opposed to using your personal computer at home, which will slow down any work required to be done on the computer. Cloud based - upload and download times Datasets tend to be considerably big in size. A smaller dataset of ~2000 images takes up ~3gb of space. This is not a significant amount of data for a local machine to transfer, but it is a considerable amount for uploading. Cost The bigger the server - the higher the rates become. Cost As opposed to a rented server - acquiring your own machine has the benefit of owning the machine, and being able to use it indefinitely (Or until it eventually breaks.) While the initial cost of acquiring an adequate machine for deep learning is higher than renting a server for a few months, it is a worthwhile long-term investment into a machine that can be used for a variety of casual or intensive tasks. Setup time Setup time Speed Speed When compared to a sophisticated server that runs many GPUs - a local machine will most likely process the training at a slower rate than a dedicated server would. A local machine will likely contain one, maybe two GPUs. Pictures are taken from the machine itself. No upload/download times. Devices Discussion After the training is done, which is usually over the span of 10's, and sometimes 100's of hours, depending on the size of the dataset and the epoch count - Running the trained model for inference only takes time in the range of milliseconds to process a single frame. R-CNN Description Disadvantages Not real-time. On average, takes 47 seconds to process a single frame. Discussion It should be noted that R-CNN has a successor called Fast R-CNN and Faster R- CNN. However, even the fastest of the successors still barely manage 5 frames a second at best. R-CNN, which stands for Region Based Convolutional Neural Networks was released in 2013. As other object detection architectures, R-CNN takes an input image, and outlines bounding boxes where it believes an item of a certain class is present. While 5 frames a second is an impressive and definitely useable result, there are alternative architectures that offer a significant improvement in inference time. Developed by Ross Girshick SSD Description SSD, which stands for Single Shot Detector. SSD was released in 2017 Developed mostly by Max deGroot and Ellis Brown Discussion Offers great framerates of an average of 45 frames per second when tested on a relatively old now graphics card, the NVIDIA GTX 1060. Disadvantages According to the Git repository, the project was seemingly abandoned about 4 years ago. According to the Git repository, the project was seemingly abandoned about 5 years ago. Discussion The component of a computer where the core computations are processed. An optional component of a computer that is dedicated and optimised in computing graphical tasks. Existing labelling related software offers quality of life features, such as rough auto labelling of the images, which only requires the user to adjust the bounding boxes and confirm their validity, rather than having to define the boxes from start to finish. Inference Example Inference Example Observations Surprisingly good results for a model trained from 120 images, with confidence values above 0.8 and sometimes over 0.9! Pre-trained model used yolov5s Architecture YOLOv5 Architecture Pre-trained model used yolov5m YOLOv5 Architecture Pre-trained model used yolov5m YOLOv5 Observations Rather poor results. Confidence values usually below 0.7, struggled to classify accurately. Observations Great results with confidence values consistently above 0.8, classifying all classes accurately! Technology utilised Deep learning computation with CPU Cores and GPU CUDA Cores running in parallel. 220Ohm resistor example Color codes Red = 2 Brown = 1 Gold = 5% tolerance 100nF capacitor example Unfortunately for the purposes of automatic identification of Integrated Circuit markings, most IC manufacturers do not follow any global standard for marking their ICs. Most manufacturers tend to have their own internal IC marking standards. Due to this fact - only known markings can be used to identify components. Mixed manufacturer ICs example This example illustrates the vast variation and lack of identifiable without access to datasheets markings. YOLOv5 Architecture A silicon board that has parts of it etched away, with only conductive tracks remaining in specific positions that are pre-planned using a CAD software. Widely used to implement electronic circuits. CAD - Computer Aided Design CAD software accelerates and automates designs in various different fields. Instructions are given to a computer that are translated into more complex and intuitive, usually GUI based interactive programs. Electrical current that oscillates. Electrical current that stays constant. An electrical component that emits light when current is passed through the circuit. A resistor that varies in resistance relatively to the amount of light the body of the component is exposed to. A ring light has been introduced for both training and inference running. Progress Issues encountered A glitch in augmentation provided by YOLOv5, where rotation during augmentation has shifted the bounding boxes of the components, causing inaccurate feedback to the model, preventing it from training appropriately. Actual bounding boxes after rotation augmentation Note the unnecessarily expanded bounding boxes. Description Submitted GitHub issue Link https://github.com/ultralytics/yolov5/issues/10639 Information gathered from replies as of todays date This issue has been reported to be part of YOLOv7 augmentation also. Example Expected bounding boxes after rotation augmentation Note the snug fit of the bounding box around the edges of the component. That is desirable, as it provides accurate information on what the model should be looking for. This will train the model in undesirable ways, detecting parts it should not. Augmentation rotation issue Software Description Augmentation Training Description Training versus Evaluation Labels Windows/Linux/Mac Desktop/Laptop Machine Discussion Discussion Specifications depend on the specific device. The specs of a desktop/laptop machine will most likely beat the specs of both a phone, and a microprocessor. Desktops are widely accessible in environments where it would be relevant to use this project, such as the home of the user, or the campus a student is in. Ease of access Ease of access Most people own a mobile device, and have it on them in most cases. Ease of access Due to the device being specialised for neural computations, it is not a common device by any means. Combined with the price tag of ~100 eur, this device will likely only be owned by developers, as opposed to users. As this device is unlikely to be owned by a user of the project, it would not be wise to require owning one to run our inference. Discussion The project will be able to support a compute stick as an alternative to a GPU. The machine must have permissions for USB connections and running the application. The app may be obtained from an App Store, that mobile devices have easy access to, as long as they have access to the internet. Specifications Specifications There are countless types of android devices on the market, all with varying specifications. Camera Platforms The application is designed through Qt Creator. Qt Creator is cross-platform. Cross-Platform Microprocessors USB computation extensions How long specifically is directly tied to the speed of hardware that the model is being ran on, and the size of the model. Even with all the speed optimisations offered by the YOLO family, a lower end device such as a Raspberry Pi 4 may take 1-2 seconds to process a single 360p image. It is important to pick appropriate hardware for your particular use cases. CPU Core Count 4 GPU Maximum Frequency 1.5GHz CPU Core Count 1 Maximum Frequency 1GHz Core Count Maximum Frequency GPU 2 532MHz CPU GPU Core Count 4 Maximum Frequency 1.479GHz Core Count Maximum Frequency 128 921MHz Observations This microprocessor is targeted towards quick graphical computations, which can instead be used for deep learning. Discussion Conclusion The original project concept at the time of project proposition had a slight shift in focus. Existing Solutions The Problem Importance of object detection Algorithmic object detection Detection of objects from an image There are various cases in which automation of object detection as opposed to having a human constantly observing footage is beneficial. Quality control Security Analysis AI based Object Detection Production In the field of security through digital cameras, object detection is an incredibly useful tool for monitoring potential intruders onto a facility. For facilities that utilise dozens, or even hundreds of cameras - object detection is a very valuable tool that requires minimal human interaction, with high level of certainty, and 24/7 attention. Depending on the security requirements, lower security facilities may not require hiring a person for constant monitoring of the security footage, and offers live notifications for any unexpected activity observed. During production of anything from farm produce, to electronic components - consistent detection and rejection of items with damage is essential. With use of object detection, flaws can be recognised incredibly efficiently, and this information may be passed onto the production line, identifying exactly which item was detected to have flaws, and be discarded automatically, without ever requiring any human interaction. If the detection is sophisiticated enough to be more reliable than a person, this opens up new opportunities for the speed and efficiency of the production, as a computer's computational power may be expanded, unlike a person. Thorough inspection of potential objects on final products is a crucial part of many fields of production. Features that may've developed during the process of manufacture may be detected on the final products. This includes positive, negative, or purely analytical features. Examples Developments of a petri dish colony PCB manufacturing error Potential signs of disease on farms Material production flaws Algorithm based detection can be used to effectively identify very specific criteria, which can be expressed as an analytical value, or a trend. Specific color based properties. Specific shapes by following line trends. Must be coherent shapes, does not perform well with partial shapes. Trees, cats, dogs, people, cars, etc. Specific patterns. Training using Deep Learning Epoch Inference A complex combination of usually many millions of digital neurons, with analog based values for each neuron, which result in a form of decision making based on a massive combination of criteria, rather than individual pixels. The process of tweaking the entire model based on the existing parameters and the current output that it produces - by use of a loss function, randomness, and clever technique, in hopes of improving the detection ability of the data the model was trained upon. These neurons work together to identify the incoming information and produce an output that resembles what the network has learned during the training of the model. Training is usually based off of a pre-trained model, that is trained on a big dataset. Most pre-trained models are trained on the COCO dataset, which is publicly available, and holds a vast amount of data. This kickstarts the model with data that it can repurpose to use as a base. Initially, the model parameters are set to a random state. The best performing model is kept between the newly trained model, and the previous best. Many epochs are ran in order to polish the model as much as possible. The output that it produces when fed input data, is of course, also random. Epochs are ran on the model to train the model. In object detection, inference is the utilisation of the model to process classifications of objects on the image. Classification The process of using the steps provided in the model to identify objects from an input image, through steps that depend on the architecture used. The identified objects are marked with a bounding box, class they belong to, and confidence in the prediction value.f Because of how far ahead the YOLO architecture is when compared to most other architectures, is utilised very commonly throughout any object detection projects. Internal steps of the You Only Look Once Inference The input image. The input image is split into a S by S grid, S being 7 in this example. Each cell predicts the bounding boxes and confidence values of each box. These steps are repeated for each of the grid cells, until everything is identified. All the identified bounding boxes. And all the probabilities for each grid cell. Each of the bounding boxes are checked for how much they cover of each probability cell, and are "shaded" in the case of this example with those probabilities. Finally, the bounding boxes are reduced using thresholds detections and NMS. NMS - Non-maximum Suppression A filtering technique used on predictions of object detectors. Picks the smallest intersection of bounding boxes, according to their confidence values. REFERENCE https://web.cs.ucdavis.edu/~yjlee/teaching/ecs289g-winter2018/ YOLO.pdf This is the final output that YOLO provides: Bounding boxes with classification (In this example, the classification is marked by a color. The real output is a string.), and the confidence value (In this example, confidence is marked by the opacity of the boxes. The real output is a number ranging from 0 to 1) The project was intended to focus on a field that is more familiar, which was more on the theme of a high precision tool, which requires a set rig to push beyond the limitations to produce high quality results. At some point, the focus has shifted towards a more flexible, but less precise concept with the consideration of more generic use through mobile deployment. As mobile deployment of object detection was not explored beforehand nor planned for, this has put additional strain on the timeline. Great results have been achieved in object detection. The model has been train on on over 3000 images that were taken and labeled over the course of several months. At this rate, a significant increase in the dataset would be required for adequate results outside of a set rig, which would require a significant amount of additional time investment that is outside of the time constraints for this project. This is of course is not a viable option in the provided time constraint, but should be considered if the project were to be continued in the future. Introduction of random rotation, with respect to the label position. Specified in a 0 to 360° range. Introduction of random spots of blur. Specified in frequency and intensity. Introduction of random scaling of the image, at a specified frequency and range of scaling. Simulates real life scenario of a different angle the feed is provided in. Simulates real life scenario of change in focus, fog, and smudges on the camera lens. Simulates real life scenario of distance. Upclose objects will cover a far bigger pixel area in an image than a far away one would, and should be trained against this to prevent detection of only certain distance away objects. Cutting and joining of images into new images. Creates additional images from the existing dataset that appear unique, making most of the existing dataset, with only a slight devalue in information stored in regards to training. Simulates bright objects interacting with the lens, ensuring the model does not get confused about glare in real life scenarios. Introduction of random glares, of specified frequency and intensity. Introduction of random spots and smudges. Simulates dust, dirt, and other particles that may be present in a real life scenario, ensuring the model does not get confused by a partial coverage of an object. Abstract Acknowledgement I am very grateful to Technological University Dublin for accepting me as a student, and providing me with the opportunity to take on this project. Declaration The material contained in this assignment is the author’s original work, except where work quoted is duly acknowledged in the text. No aspect of this assignment has been previously submitted in any other unit or course. IDE - Integrated Development Environment A sophisticated text editor, designed specifically for code development in certain languages. Most IDEs today come with a high range of optional plugins that can be used to further increase production speed, and reduce redundant tasks via automation of said tasks. A term often used in deep learning, which is the process of attempting to simulate intelligence that is similar to that of a human, by the use of a computer, in order to tackle issues that a standard style of computer operation either fails to, or performs at incredibly slow rates. Course TU807 Year 4th Title Bachelor of Engineering (Honours) in Computer Engineering in Mobile Systems Code I’d like to acknowledge and express my gratitude towards my project supervisor, Benjamin Toland – who has taken on me and my custom project and has provided excellent guidance and feedback throughout the development of this project. I am also very grateful for the existence and availability of search engines, the primarily used for this project being Google. It is a valuable resource for research, even though you should not take every search result at face value, and ensure at least a few trusted sources agree with the findings. I would also like to thank my wonderful old and new friends and peers that I was able to meet thanks to my ability to attend TUD Blanchardstown Campus, and the quiet spaces provided for us to further our education with minimal interruptions. I am also grateful to my brother, Justas Bartnykas – who is also an engineer. I am grateful to him for introducing me to Qt Creator, which is now my go-to IDE for developement of C++ code for the past over 5 years now. I have also been provided access to a fairly expensive camera that he owned, that allowed the project to begin sooner than it otherwise could've. I’d like to offer my sincere apology and thanks to anyone else I may have missed that has contributed to this project in any way. When we look at an image, we can immediately discern objects that are displayed, without ever having to think about it. It is effortless. We use a combination of past experiences and deduction to determine information almost immediately. This ability of ours as humans is thanks to millions of years of evolution, in a world where not detecting a potential threat makes a difference between life and death. It is in our subconscious nature to detect objects within a moment's notice. The problem arises when we want to implement a computer to process object detection for us. A computer is built upon the most simple of arithmetic tasks, combined together from an incredible amount of transistors that perform these tasks, into massive systems - which operate entirely in digital binary. A computer has no concept of learning from past experiences, nor any feelings that may affect its decisions. Given the same instructions, it will produce the same results on its first day of manufacture, and the last day of its operation (provided the unit was not damaged in a way that it would yield unpredictable results). Computers are designed to process computational tasks at complete precision and consistency. When a computer is exposed to an image in form of a pixel grid, all it truly sees is numerical values that are assigned to each cell of the grid. Computers have no concept of color, let alone real life objects. The simple action of either moving or scaling an object completely changes the arrangement of the pixels that represent this object. To a person - this is no issue. It is obvious that this is still the same object. However, to a computer - the data has just shifted around entirely. Examples A close up of a 220ohm resistor An even further close up of the first band, where each pixel is beginning to show The same band, rotated at an arbitrary angle The width and height of the image have both changed due to this rotation. The order of the pixels has changed drastically. However, it will struggle heavily at detecting and identifying any generic object. Examples Examples While this is a complex issue to tackle already, it only becomes more complex when we want to detect something generic, such as a tree. A tree has many properties that may differ, such as: presence, color, shape, and size of leaves/needles, thickness, size, and color of trunk, branching styles, etc. While still being effortlessly identifiable by an intelligent creature. The dataset is separated into two categories - one on which the model is being adjusted on, and another on which the model is being ran to determine the new confidence values achieved by the alterations. Both datasets must be labeled for fully automated evaluation. Labeling is an essential part of training for Object Detection. Labels are bounding boxes that determine what object is present where in the input dataset, in order for the model to do its best to detect them. Neural Networks The networks are trained through deep learning. Neural networks are an imitation of biological creatures intelligence, which is known as Artificial Intelligence. AI is designed to be ran on a computer. Description It has the ability to be trained, which simulates past experience of biological intelligence. In the scope of this project, somewhere between 250 and 500 epochs will suffice. The boost in results is limited to the set rig. Author B00125142 Violet Concordia Supervisor Benjamin Toland Several of same type of component, all packaged in a single line. Example: an SIP resistor, which features multiple resistors, all connected to a single ground pin. Observations Calculations Calculations Gantt Chart as Table Choice Camera Rig Ring Light Model OSY-C100-4-12 Specifications HDMI USB Framerate 30fps Video Image MegaPixels Pixel Size 1.55µm × 1.55µm 12 4k 1080p Framerate 60fps 1080p Framerate 30fps Chosen mode of operation USB connection, 1080p, 30fps. This mode was chosen due to the high comptability range of a USB connection, and both the resolution and framerate being more than sufficient for the purposes of the project. Specifications Connection AC power plug. Dimensions Height Width Length 22.6mm 95mm 59mm 65mm is the maximum diameter this ring light can be attached to. Features Adjustable intensity via a knob on the side. There is a high variety of camera rigs available for purchase and for the purposes of this project - as long as the rig has a surface and the camera has a convenient way to be mounted - there is not much of a difference on which rig is chosen, provided it is adequate in dimensions and is capable of steadily mounting the camera. That is the reasoning for choosing a rig that is quite basic, but does the job well, and was already available at home. Dimensions Base Width Base Length 280mm 180mm Choice Rationale Choice rationale For the purposes of object detection at the scale of this project, 1080p at 30 frames per second is plenty of incoming data. The quality of the output images has been tested to be sufficient. While this camera is over-qualified for the project, as mentioned in the Acknowledgement section, this camera was available at the start of the project, and was chosen out of convenience of availability. Height 175mm to 350mm (Adjustable arms) Choice Rationale This ring right was chosen due to being far more than sufficient in supplying adequate lighting, and being reasonably cheap. Cost ~250€ Cost ~15€ Cost ~40€ This report illustrates the development of a 4th year Computer Engineering Project, which is designed to tackle the issue of computer-based detection of objects present in any given image from a camera image/video feed. While object detection is a very vast topic, the project focuses specifically on electronic component detection and identification. Specialisation to read additional information from the electronic components, such as their rating values, or marking codes. The focus audience of this project is engineers. Hopes to assist people in becoming engineers by providing a useful tool for identification, as well as curiosity to investigate both the electronic components, and inner workings of the project. It also aims to be a valuable tool for existing engineers by assisting identification of electronic components in bulk. The project utilises Artificial Intelligence to achieve its goals, with high confidence values. To run AI, a model is required. The model must be trained from a dataset. The dataset for the training of the model is gathered through the set rig, which is a camera stand with a surface that the items are placed on, and the camera with a ring light positioned directly above the placed items, for optimal and consistent viewing angles and visibility. The project is controlled through a GUI application that provides feedback on the information gathered, and tools to interact with the device. Currently the GUI application is deployed on a desktop machine, and the camera on a set rig. The desktop receives the information from the camera through a USB connection. The purpose of the set rig is to both provide an efficient environment for data gathering purposes, and to constrain the amount of variation a tested-upon image has exposure to, which drastically reduces the angles, lighting conditions, distance, and background variations of the input images, which is highly beneficial for a higher confidence value of the object detections. For the ideal future which may be out of scope of the current time and resource constraints Thanks to the project being developed and running on C++, in addition to using the YOLO architecture – the project has the opportunity to expand into a high confidence, high efficiency, constantly-improving through user-submitted dataset project, that is deployable on desktop computers, laptops, microprocessors, and mobile (Android/iOS) devices. Live Labeling The ability to take a snapshot of the current frame, defining appropriate labels, and saving this labeled snapshot for future training. All from inside the GUI. Alternatively, taking snapshots of the GUI and saving them for later labeling. It should be noted that the labels must be adequately labeled. If mass deployed with the ability to live label and submit the images to a server for further training, there must be a sophisticated method of ensuring that the submitted data is 100% accurate. If an image is mis-labeled, or an object is missed – this would cause mis-training, or worse – de-training of the model, corrupting the dataset and setting the training backwards rather than forwards. Potential solutions Manual inspection of submitted images, for the purpose of quality control, which is crucial for the success of the model. Submission of Labeled images Requires manual inspection. Un-labeled images Requires manual labeling. Both of the solutions pose an issue of requiring manual labor from a person that is not the user, and is adequately knowledgable of which data is sufficient both in image quality, and labeling – which causes a linear demand of paid workers for the submission flow rate. In conclusion - while the project did not meet all the new goals - it did explore them in detail. The training process for the model was well-planned and executed, resulting in a well-tuned model that is capable of meeting the criteria for a set-rig operation of the project. Setup OpenCV Extra Modules Qt Creator CMake Version 3.16.3-1ubuntu1.20.04.1 Description A free, open-source software, designed to automate building, testing, packaging, and installing software, by using compiler-independent methods. Platform Cross-platform CMake does not build - rather, it generates another system's build files. The user is able to specify exactly which components they would like to be included, and their properties. The generated build files are then used to build from. Required items Installation Linux Terminal command to install cmake-gui OpenCV CUDA Download the appropriate version for your machine https://developer.nvidia.com/cuda-downloads? target_os=Linux&target_arch=x86_64&Distribution=Debian&target_version=11&target_typ e=deb_local From a terminal in ~/opencv/ Let ~/opencv/ be our cmake directory git clone https://github.com/opencv/opencv.git git checkout 4.7.0 From a terminal in ~/opencv/ git clone https://github.com/opencv/opencv_contrib.git git checkout 4.7.0 cd ./opencv cd ./opencv_contrib Clone the OpenCV source repository Enter the repository directory Checkout this specific version Clone the OpenCV Extra Modules source repository Enter the repository directory Checkout this specific version Download Link Extract into ~/opencv/[filename] This command may be used: tar -xvf ./cudnn-linux-x86_64-8.6.0.163_cuda11-archive.tar.xz Note: v argument is optional. V stands for verbose, which indicates to print out every file being extracted. Finally, the ~/opencv/ directory should have the following content Working directory We also need an empty ~/opencv/build directory for the output of CMake Note: Libraries Generation of build files using CMake GUI Open CMake enter cmake-gui into the terminal First we have to select the source and binaries paths Source path should be ~/opencv/opencv binaries path should be ~/opencv/build Then click configure Configure is highlighted in blue This setup has been carried out on the previously mentioned Lubuntu 20.04 machine. Steps may slightly vary on other Operating Systems. A custom UI theme was used on these machines. Press Configure OpenCV Neural network Qt Creator Extras If any of the options were not available, simply press configure again and they should appear. Press Generate. ENABLE_FAST_MATH CUDA_FAST_MATH OPENCV_ENABLE_NONFREE WITH_PNG WITH_FFMPEG BUILD_opencv_dnn Qt6_DIR OPENCV_EXTRA_MODULES_PATH WITH_GTK_2_X WITH_GTK Qt6* WITH_QT WITH_ONNX CUDA The default options of Unix Makefiles and Use default native compilers will most likely suffice. Choose Finish. Missing directories (follow pattern of others) /opt/Qt/6.2.4/gcc_64/lib/cmake/Qt6 WITH_NGRAPH WITH_INF_ENGINE WITH_CUDNN WITH_CUDA CUDA_TOOLKIT_INCLUDE CUDA_VERSION CUDA_TOOLKIT_ROOT_DIR CUDNN_INCLUDE_DIR CUDNN_LIBRARY OPENCV_DNN_CUDA ~/opencv/opencv_contrib-4.x/modules Make sure expected version is found. 11.4 was chosen for this project. ~/opencv/cudnn-linux-x86_64-8.6.0.163_cuda11-archive/lib/libcudnn.so ~/opencv/cudnn-linux-x86_64-8.6.0.163_cuda11-archive/include /usr/local/cuda/include /usr/local/cuda Flags setup The reason for this is that some flags cause other flags to appear. For example: WITH_CUDNN is required for CUDNN related settings. Repeat until all the flags are as specified. The ~/opencv/build directory will be populated with all the necessary build files. Build Navigate to ~/opencv/build from your terminal nproc Execute the comand Returns how many cores the machine has From ~/opencv/build, execute the command make -j $(nproc) The command then becomes make -j 8, if the machine had 8 cores. Installation sudo make install Throughout the project, the machines presented were not consistent. A personal laptop, and a personal desktop were used. Performance-wise, this is insignificant, as both machines are running the same version of Lubuntu 20.04, and both utilise a NVidia GPU that has CUDA cores. This process took somewhere around 4 hours. The duration of this process is heavily dependent on the specs of the machine. OpenCV Finally, OpenCV is now fully installed, and ready to be used. Project Qt Widgets Application Project name and path options Build System: CMake Build Kits For the dev environment and basic setup, we want to use the Desktop kit. In the future, an Android kit may be necessary, to deploy the project on Android devices. We are using CMake to generate build files once again, this time the terminal version. Creation Setup CMakeLists.txt (project file) configuration - including the required libraries # CUDA_START set (CUDA_TOOLKIT_ROOT_DIR "/usr/local/cuda-11.4" ) find_package (CUDA 11.4 REQUIRED) set ( CMAKE_CUDA_STANDARD 11) set ( CMAKE_CUDA_STANDARD_REQUIRED ON ) # CUDA_END #OPENCV_START find_package (OpenCV REQUIRED) include_directories ( ${ OpenCV_INCLUDE_DIRS } ) include_directories (4th_year_project PRIVATE "/usr/local/include/opencv4" ) target_link_libraries ( 4th_year_project PRIVATE ${OpenCV_LIBS} ) #OpenCV_END The CUDA version matches the one compiled during OpenCV installation Installation Version 6.2.4 The Open Source version was used for this project. We want the Qt Online Installer. The system is running Linux - we need the matching version of the installer. Online Installer - version selection The UI design overview Note: The design of the project is up to interpretation, and only key parts will be covered briefly. This project is open source, and the reader may clone the repository to run it from this point on. The rest of this section will no longer be a tutorial, but an overview of the project internals. Systems Inference Post-processing Camera UI Orientation Optical Character Recognition Confidence Threshold Score Threshold NMS Threshold Detection Struct classId Internal logic, does not concern the user - used to determine the class name. className Name of the class (Resistor/Capacitor/etc...) extra Post processing information, appended to the name during display confidence How confident the inference is that this detection is indeed this class, and right here. Ranges from 0 to 1, 0% to 100%. color Display color of this detection. box The location and size of the detection, represented as a box. mat A cropped out image from the original, from the position and size of the box. Confidence requirement, under which the detection is rejected. Score requirement, under which the detection is rejected. NMS requirement, under which the detection is rejected. Saturation adjustment Utilises Histogram equalisation provided by OpenCV libraries to roughly adjust every detection to a specific saturation value, in order to make post-processing more consistent. Draws countours around a greyscale version of the component, in order to determine orientation. The images are automatically rotated to face forward. The extra space that is caused by rotation is filled with pure black pixels. Resistor Decoding Algorithm Optimisation The decoding algorithm method The algorithm was designed on a separate project, and the values that were found to do the job best were left as defaults. Arguments inputMat The cropped out, prepared, and rotated input image of a resistor. Example results Final result of the algorithm, identifying 4 bands with colors: Yellow, Purple, Green, and Gold. Bulk calibration Early bulk testing Determined resistor values are displayed as the last numbers in the window name. The algorithm achieved 100% success rate for this particular random batch test. Bulk example Before saturation adjustment After saturation adjustment Base algorithm testing Link https://www.qt.io/download Boundaries Thanks to the set shape of a resistor, the algorithm is able to narrow down on which area the bands are present in. This means there is no need to process outside of the predicted magenta bounds. This reduces the processing power required to an average of 1/4th (difference between the entire image area, and the insides of the magenta bounds.) Glare destroyed information filtering qreal middleCropBoundPos Ranges 0 to 1 Controls how much of the vertical middle is cropped out. (Will be covered in further detail under the Glare section.) qreal horizontalBoundPos Ranges 0 to 1 Controls the horizontal magenta bounds position qreal verticalBoundPos Ranges 0 to 1 Controls the vertical magenta bounds position HSV Filter Minimum/Maximum Hue Saturation Value Filters pixels inclusively/exclusively (depends if min or max is the higher value), replacing them with 0, 0, 0 rgb (black) bool display True or False Debug feature, visualises the internal workings of the algorithm. Used to provide the media in this report. Ranges 0 to 1 Ranges 0 to 1 Ranges 0 to 1 QString displayName The name of the debug display, if it is enabled. Overexaggerated vertical crop from the middle middle, for demonstration purposes Since the information is destroyed by glare, there is no reason to process the white strip in the middle. Cropping out from the vertical middle leaves us with more information in the line horizontally. Result after HSV filter Steps Further preparation processing The algorithm was originally designed on a separate work environment, with sliders to control every key parameter. Dataset Over 1500 pictures of resistors were gathered from the rig itself, for the algorithm to be evaluated against. Resistor body and background filtering, using the HSV filter arguments parameters. Before HSV filter After HSV filter The band extraction algorithm Steps Slight blurring to average out the bands colors Before blur After blur Example Before optimisation After optimisation Note: The algorithm is calibrated to the optimised version. It expects a specific ratio between color and blank pixels in the entire row that is being tested. Notice the lack of processing outside of the bounds Average of every row's hue and count of non black (empty) pixels are obtained. Initialisation For every row top to bottom, ratio of non-zero pixels is compared to the width of the row being processed. If the count is above a certain treshold, the algorithm assumes this is a color band, and enters band mode. Once in band mode, the algorithm starts to keep track of the average hue of the band. Otherwise, it continues to the next row, until the condition is met, or no more rows remain. This is what determines the final color of the band. Once the algorithm detects an insufficient ratio of color in the row, band mode ends. If the tracked band height was more than a certain small threshold (Aims to prevent noise counting as bands), pushes the average hue onto the Bands List, and advances a certain step size forward, because there will be no band immediately after a band. The steps loop, until no more rows remain. Color determination The final band colors identifier names are determined from a combination of the Hue, Saturation, and Value of the colors. For example, red/brown can only be within ~330-360 and ~0-30 range. This can be further processed by analysing the saturation and value of the color to determine if this was a red, or brown color. Connection maintenance Frame retrieval The application starts with the connected flag being false. This flag determines if it should try connecting to the camera. While this flag is false, a QTimer will run, which is a separate thread that executes every x amount of time. This timer is responsible for attempting to connect to the camera, until it is connected. If we are unable to communicate to the camera, this indicates it was disconnected, and the connected flag is set to false again. Once the connection is established, the connected flag is set to true, and the timer ends. To replace the connection timer, a frame retrieval timer starts in its place. The connection is maintained completely automatically. A frame is requested from the camera every specified amount of time, depending on the frame rate. Each frame is sent off to processing, while another one is being retrieved asynchronously, thanks to QTimer threads. Zoom control For ease of focusing the camera and inspection of components upclose. Status bar for configuring what gets ran/displayed Comes with a reset position to middle button. Inference detections Post processing Export button Exports the entire frame, both before and after inference. Each detection is exported separately. Live camera feed Console and Inference/Post-processing information output boxes Provide general information about the detections. This can be achieved by navigating to an appropriate directory to hold the repository, and entering the command: git clone https://github.com/Harmonised7/ai_identifier.git If you are using SSH instead of HPPS, the command is instead: git clone git@github.com:Harmonised7/ai_identifier.git This is the reason the gold band was not found in the unoptimised version. Found optimal calibration 70% horizontal cut off (vertical magenta bounds) 25% vertical cut off (horizontal magenta bounds) 12.5% cut out from the vertical middle Full Hue range 35% to 100% Saturation range Full Value range The brightness of the images now matches the others to a considerable degree. This provides us with a more consistent workspace for the algorithm to work from. Recreating these features may be considered recreating the wheel. Threads Simplified overview of the utilised threads. Note that the frame updating thread utilises both CPU and GPU threads while running inference and post-processing. Some of the steps may appear slightly different on your machine. Overall, YOLOv8 at the time of writing was not at a state in which it would be preferable to utilise it in favor of YOLOv5. YOLOv8 is still in developement and while the planned features will most certainly be incredible - YOLOv5 was chosen due to reliability and availability of the existing features. Android deployment Qt Maintenance Tool Android Studio Setup a device Qt Creator Install the correct Android SDK Setup kit Android related libraries selected Android Ninja Overall - all the initial criterias and some beyond have been achieved with great success. However, post-processing came at a cost of a considerable amount of processing power - has yielded very impressive results that could've only been hoped for. A further boost in performance was instead achieved by transitioning to a bigger model. This of course came at a cost performance, dropping from camera side hardware limitation, to the laptop that the project was developed on limitation. UI improvements Contrast text Post Processing Post processing identifying colors Horizontal text adjustment Horizontal text adjustment: middle Horizontal text adjustment: right side Horizontal text adjustment: left side Examples The text position adjustment was designed to always fit text on the screen, without it leaving the screen. To achieve this - the horizontal width of the screen and box, position of the box on the screen, and the display text width had to be considered. It is designed for the text to start directly from the start of the component when the component is all the way on the left, and to end at the end of the component when the component is all the way to the right, with a smooth transition in-between. The text creation was upgraded from plain color text, to a bright outline with a darker color fill. This resulted in ease of readability of the text under all background conditions. Examples Plain text Contrast text Used in precise inspection of the components and camera focus adjustment. Examples Zoom example 3 Zoom example 1 Zoom example 2 No zoom The project was an overall great success both as a learning experience, and as a utility tool it has set out to be. Unfortunately, it has been concluded that even if the dataset has been expanded a few times - mobile deployment with a live video feed would simply not be feasible due to the lack of CUDA core availability on mobile devices, and the lack of a set rig environment that the project has been designed for. With more resources and advances in mobile technology - this project has the potential to one day make it into the mobile device market. An option that was heavily considered was to move the processing part onto the cloud to compensate for the lack of computational power, however - the project's core design was to provide live video feed processing as an assistant tool the user may use for live reference, as they move the components around. Reduction from live video on a screen, to manually taking pictures and waiting for the results to come back would've interefered with the core idea to a point where it would be a different project, but it was very insightful to research the options and has inspired several future projects. The post processing algorithms were carried out as initially conceptualised, and turned out to be a great success! The initial goals of the project were fulfilled with better than expected results for the timeframe that was available. Even in the most ideal of conditions - OCR has more often than not failed to provide adequate results, while costing a significant amount of processing power. OCR has proved itself to be rather unreliable. Furthermore - the testing was done while facing the component upright. While this is certainly achievable, it is a further hit to the processing power, because each rotation of 0, 90, 180, and 270 degrees would require a test - increasing the total required processing power to over 4 times of what simply running OCR on a prepared image would cost. It was been concluded that OCR is simply not something that is feasible to be utilised during runtime for this project. Perhaps with further developement in OCR efficiency, the project may expand to support OCR in the future. Some of the components may have their surface worn, while others are very reflective and cause significant glare. In a real life scenraio, the component will face any random direction, so it must first be rotated in the exact direction to face upwards to be readable. Cloud Machine Specifications Performance directly correlates to the cost. Discussion Workload The inference and post-processing workload would be processed solely on the cloud. New constraints Having a live, lossless video feed would be essential for adequate results. This would put a huge strain on the required processing power on the user's device. A compromise could be to replace live video feed processing with image submissions instead. However, even a mobile device would likely be able to complete processing a single frame in the duration it would take to send the data on to the cloud, the cloud to process it, and then receive the processed data back. This approach in the end may not reduce the processing power significantly enough, and may actually introduce latency, depending on the connection qualitty. Because the processing would be done on the cloud, access to internet would required. Cost Running a server costs a monthly fee that may be avoided by running the server from home, but doing so would hinder the expandability of the project, most likely resulting in dissatisfaction through latency for the users. Post processing identifying ohm values